BioMCP
BioMCP is an MCP server for federated bioinformatics and biopharmaceutical database access, with optional local database and sandboxed analysis toolboxes.
Search and retrieve genes, associated diseases, drugs, trials, articles, and Reactome pathway enrichment.
Search and inspect variants, including OncoKB cancer annotations and variant-linked trials.
Search drugs and diseases, and find drugs or trials associated with them.
Federated literature search across PubMed, EuropePMC, Semantic Scholar, arXiv, PubTator, and LitSense; retrieve articles, preprint full text, and journal metrics.
Search and retrieve clinical trials, PDB structures, patents, GEO, SRA, GenBank/RefSeq, GTEx expression/eQTLs, and Ensembl annotations, homology, VEP consequences, and genomic regions.
Resolve concepts with
discoverand retrieve multiple entities in parallel withbatch_get.Optionally query MySQL/SQLite databases, run Bioconductor DESeq2/edgeR/limma in WebAssembly R, and run samtools/bedtools/bcftools analysis in WebAssembly.
Inspect and configure optional features via
biomcp_configure; environment-only secrets remain query-only and masked.
Provides access to biomedical literature through PubMed, enabling search and retrieval of articles, and citation data via the article_search and article_get tools.
Integrates with Semantic Scholar for federated literature search and citation analysis, supporting article search and citation graph data.
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@BioMCP get info on BRCA1 gene"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
BioMCP

Highlights
Democratizing agentic access to bioinformatics and biopharmaceutical databases and analyses.
Section-based federated access to 50+ bioinformatics, pharmaceutical, and patent databases
Optional toolboxes for local database curation and dependency-free analysis with Bioconductor and SAM/BED/BCFtools — no R installation, C toolchain, or containers
Concrete example vignettes, developed fully in the open
Related MCP server: PubMed Advanced MCP Server
Install
npx -y biomcp doctor # diagnose a machine: Node gate, config health, feature gates, peer deps
npx biomcp # zero-config stdio MCP server (this is what MCP clients run); Node >= 22.13Setup is guided in docs/AGENT-INSTALL.md — a one-minute start, copy-paste config entries for Claude Desktop, Claude Code, Codex, and OpenCode (one canonical pinned command covering every feature), biomcp doctor as the single troubleshooting entry point, and agent-friendly paths for API keys and optional features.
Multiple agents on one machine: biomcp enforces arXiv's Terms-of-Use rate limit (1 request / 3 s) per process. When running multiple biomcp MCP instances on one machine (e.g., one stdio instance per agent), rate limits are not coordinated across processes, and arXiv's TOU applies to all your machines in aggregate. For multi-agent single-machine deployments, run a single shared biomcp serve daemon (see docs/SELF-HOSTING.md) so all agents share one process and one global rate limiter.
Available Tools
Full tool schemas (params, enums, defaults) live in src/server/README.md.
Gene (7)
Tool | Description |
| Search genes by symbol, name, or keyword with chromosome filter |
| Get detailed gene info by HGNC symbol with optional sections (core, pathways, protein, ontology, go, interactions, expression, protein_atlas, constraint, druggability, dosage_sensitivity, clinical_evidence, disease_associations, diseases, funding). Set |
| Get diseases associated with a gene (DisGeNET / OpenTargets) |
| Find drugs targeting a gene (OpenTargets; envelope |
| Find clinical trials for a gene |
| Find articles about a gene (federated across PubMed, EuropePMC, Semantic Scholar, arXiv, PubTator, LitSense; envelope |
| Pathway enrichment analysis for a gene list (Reactome) |
Variant (4)
Tool | Description |
| Search variants by rsid, HGVS, gene, ClinVar significance, frequency, CADD |
| Get detailed variant info with optional sections (frequency, predictions, clinical; |
| Get OncoKB cancer variant annotations (requires |
| Find clinical trials for a variant |
Drug (3)
Tool | Description |
| Search drugs by name, mechanism, or keyword (envelope |
| Get detailed drug info with optional sections (us_regulatory, who_regulatory, safety, targets, indications, adverse_events — FDA FAERS reactions ranked by report count; WHO PQ listing lookup, ~650 global-health priority products) |
| Find clinical trials for a drug |
Disease (4)
Tool | Description |
| Search diseases by name, phenotype, or keyword (envelope |
| Get detailed disease info by ID (DOID, MONDO, OMIM, etc.) with optional sections (gene_associations, phenotypes, pathways) |
| Get drugs for a disease (OpenTargets; envelope |
| Get clinical trials for a disease (ClinicalTrials.gov; envelope |
Article (4)
Tool | Description |
| Federated literature search across PubMed, EuropePMC, Semantic Scholar, arXiv, PubTator, and LitSense with optional date range filtering; |
| Get detailed article info by identifier (PMID, PMCID, DOI, or bioRxiv/medRxiv preprint DOI) with optional sections: |
| On-demand bioRxiv/medRxiv preprint full text (JATS XML parsed to sections), by preprint DOI with optional version (default: latest). Strictly on-demand (never during search), downloads are paced (≥3 s spacing), cached per (doi, version) — JATS per version is immutable — and license-aware ( |
| OpenAlex journal quality metrics (2yr mean citedness — an open-access impact-factor alternative), h-index, quality lists; by name, NLM abbreviation, or ISSN |
Trial (2)
Tool | Description |
| Search clinical trials by condition, intervention, status, or phase. Cursor-based pagination via |
| Get detailed trial info by NCT ID with optional sections (eligibility, locations, outcomes) |
Utility (2)
Tool | Description |
| Free-text concept resolution across all entity types |
| Retrieve multiple entities in parallel |
Structural Biology (1)
Tool | Description |
| Search PDB structures, get entry metadata with optional sections (polymer entities, ligands, assembly, experiment, citation), and download structure files (mmCIF/PDB) |
Patents (2)
Tool | Description |
| Search patents worldwide (US, EP, WO, JP, 100+ authorities) with assignee/inventor/CPC/status/date filters and relevance ranking ( |
| Get patent details by publication number with sections: abstract, claims (US fulltext via USPTO Public Search; EP/WO via EPO OPS when configured, otherwise Google Patents pages), citations (forward + backward), family, classifications |
GEO (2)
Tool | Description |
| Search NCBI GEO for functional genomics studies (expression microarrays, RNA-seq, single-cell series) by entry type (GSE/GSM/GPL/GDS) and organism; results carry cross-links (bioproject, pubmed_ids; sra_project when NCBI links the SRA study — otherwise geo_get's sra field) for chaining |
| Get the full SOFT record for a GEO series/sample/platform: summary, organisms, sample preview (≤20), supplementary file URLs, and cross-references; optionally download the first supplementary file. GDS curated DataSets return structured guidance ( |
SRA (2)
Tool | Description |
| Search NCBI's Sequence Read Archive for sequencing experiments and runs by free text, accession, or field syntax; returns experiment/study/sample accessions with library strategy and run counts |
| Get full details for an SRA accession: run (SRR/ERR/DRR — instrument, spots, bases, size), experiment (SRX/ERX/DRX — library design), study (SRP/ERP/DRP — experiment list), or sample (SRS/ERS/DRS); NCBI SRA indexes NCBI, ENA, and DDBJ accessions alike |
GenBank (3)
Tool | Description |
| Search NCBI nucleotide records (GenBank/RefSeq/INSDC) by plain terms, accession, or field syntax; results include accession.version, definition, length, organism, topology |
| Fetch a GenBank/RefSeq record as GenBank flat file or FASTA; whole records capped at 2 Mb — larger records require a |
| Map a GenBank/RefSeq accession to its NCBI Gene IDs (elink nuccore→gene) for bridging into gene tools |
GTEx (2)
Tool | Description |
| Get median gene expression across GTEx tissues (Analysis v10, 54 tissue sites, TPM, highest first); accepts HGNC symbol or Ensembl gene ID, with optional single-tissue filter |
| Get significant cis-eQTL associations for a gene in a specific GTEx tissue (v10): variant_id, p_value, NES, slope, sorted by ascending p-value |
Ensembl (4)
Tool | Description |
| Resolve a gene in Ensembl terms for any of ~356 species: stable ID (+version), symbol, coordinates on the current assembly, canonical transcript; |
| Find orthologues/paralogues across species via Ensembl Compara — target stable IDs, taxonomy level, percent identity, sorted by identity; filter with |
| Compute variant consequences on demand via Ensembl VEP for NOVEL variants and non-human species: most severe consequence, per-transcript effects (SIFT/PolyPhen), co-located ClinVar/COSMIC/gnomAD data. Known human variants get deeper pre-computed scores via |
| Query genes/transcripts/known variants in a genomic interval ( |
R Analysis (4, optional — ANALYSIS_R=1)
Tool | Description |
| Differential expression for RNA-seq counts with Bioconductor DESeq2 (negative binomial, independent filtering, optional LFC shrinkage) in sandboxed WebAssembly R. Inputs: integer count matrix + sample metadata + design formula; output: markdown table of top genes by adjusted p-value with summary ( |
| Differential expression with edgeR — TMM normalization, empirical-Bayes dispersion, quasi-likelihood F-test ( |
| Differential expression with limma-voom — precision-weighted linear models with empirical-Bayes moderation; same input/output contract |
| R runtime report: R/webR versions, installed package versions, memory, mirror endpoint — for diagnosing analysis issues |
First use starts a ~1 GB WebAssembly R worker and downloads the wasm package bundle (~62 MB) from GitHub releases (cached). Requires webr installed next to biomcp. Guide: docs/R-ANALYSIS.md.
Biowasm Analysis (8, optional — ANALYSIS_BIOWASM=1)
Tool | Description |
| Inspect an alignment (SAM/BAM/CRAM): header contigs, sample/read groups, flagstat mapping metrics, per-contig counts via idxstats when indexed — "what's in this BAM?" before region work |
| Reads, depth, pileup, or read extraction in a genomic region (samtools view/depth/mpileup); indexed sources use fast positional retrieval, indexless sources stream a BED filter (depth requires coordinate-sorted input and detects order violations), returning counts, coverage tables, SAM rows, or a BAM artifact |
| Inspect a VCF/BCF: contigs, sample count and names, INFO/FORMAT field inventory from the header |
| Variants in a region as a narrow field projection (bcftools query): chosen columns, sample subsets, expression filters, variant types — or a sliced VCF.gz artifact |
| Interval algebra on BED tracks (bedtools intersect/merge/subtract/coverage/jaccard/sort) with the streaming |
| Format plumbing: SAM/BAM/CRAM via samtools view, VCF/BCF via bcftools view, VCF/BCF → TSV via bcftools query; results are artifact handles reusable as |
| Biowasm runtime report: pinned tool versions, asset cache state, engine status, retained artifacts, memory |
| Constrained escape hatch: an allowlisted samtools/bedtools/bcftools subcommand with schema-validated args (no shell, paths under /shared only) |
First use downloads checksum-verified wasm assets (~4.5 MB, cached); no extra npm packages. Indexed sources answer region queries with fast positional retrieval (~0.2 % of file read); indexless sources fall back to streaming BED filters. Guide: docs/BIOWASM-ANALYSIS.md.
Citation Module
Citations federate 5 providers in fast (~4s) or full (~15-30s) mode. Forward citation lists come from Europe PMC, OpenCitations, and Semantic Scholar; Crossref supplies counts and backward references. Provider matrix and schema details: src/server/README.md.
Optional Features
Capabilities that ship with the package but stay inactive until enabled. Each links to its own guide:
Feature | Enable | Guide |
Database access — read-only SQL tools ( | Set | |
R analysis — Bioconductor differential expression ( | Set | |
Biowasm analysis — samtools/bedtools/bcftools (BAM/BED/VCF) in sandboxed WebAssembly; streams/indexes real human-scale datasets (~300 MB BAM scans, region queries touch ~0.2 % of the file); assets ~4.5 MB cached at first use; no extra npm packages | Set |
Instead of hand-editing env blocks, agents (and users) can self-serve through the always-available biomcp_configure tool: it reports every parameter's status/provenance, writes the .biomcp.json project config file for the optional features above (env vars keep precedence; env-only parameters are query-only and value-masked), validates changes, detects conflicts, checks peer-dependency prerequisites, and spells out the restart/verify steps. Details: docs/ENV-VARS.md → Project config file.
Documentation
Doc | Contents |
Guided installation & client configuration (Claude Desktop, Claude Code, Codex, OpenCode) | |
Single source of truth for every environment variable | |
Database access feature guide | |
R analysis feature guide (Bioconductor in WebAssembly) | |
Biowasm analysis feature guide (samtools/bedtools/bcftools in WebAssembly) | |
Build, test, publish workflow | |
CI pipeline, Dependabot automation, auto-merge safety model | |
Full tool schemas (params, enums, defaults) | |
User-agent E2E tests for the analysis tools |
License
Licensed under the Apache License, Version 2.0. See NOTICE for attributions.
BioMCP-TS is adapted from the upstream BioMCP Rust project (MIT) with an agent-first development approach and enhancements — kudos to the original authors.
Available Tools
43 toolsarticle_getARead-only
Get article by PMID/PMCID/DOI. Section response keys differ from request names: "oa" is stored as sections.open_access, "graph" as sections.citation_graph. Preprint DOIs (bioRxiv/medRxiv, e.g. "10.1101/2025.03.05.641768" or new prefix "10.64898/...") return the preprint record: all versions, license, category, funders, JATS full-text URL, published-version mapping, and (with sections=["citation"]) Europe PMC preprint citations. Journal articles, sections=["fulltext"]: returns the ACTUAL full text when freely readable in PMC/Europe PMC (open-access AND free-to-read). fulltext.status is "available" or "unavailable"; when unavailable, fulltext.reason is one of paywalled | no_pmcid | no_machine_readable_xml | not_implemented_for_preprints | fetch_failed (with alternatives in links, e.g. doi_url); a transiently failed OA-status lookup counts as no status info, so a paywalled in-PMC article surfaces as fetch_failed until a retry reclassifies it as paywalled. The five-value open-access vocabulary — open_access | free_to_read | paywalled | not_in_pmc | unknown — is sections.open_access.status (request sections:["oa"]); fulltext.oa_status echoes it so paywalled vs not-in-PMC is visible without the oa section. Default (fulltext_output="file"): the complete text is written to a server-side markdown file and the reply carries file_path plus an outline with per-section line numbers and char counts — Read/Grep the file selectively instead of loading it into context. fulltext_output="inline" (use only if you cannot read server-side files): fulltext.sections is an ordered array of {heading, text, truncated} objects plus char_count/total_char_count/truncated/omitted_sections, capped by fulltext_max_chars. sections=["oa"] returns license/access metadata and links ONLY (pdf_url presence does not by itself mean the text is fetchable — use fulltext for content). "all" = oa+annotations+graph+citation WITHOUT fulltext (request ["fulltext"] separately). oa/fulltext do not apply to preprints (see the jatsxml_url link). Journal-article citation: fast mode (~4s, 4 providers, auto-fallback to PubMed) or full mode (~15-30s, all 5 providers incl. PubMed). Forward citation lists come from Europe PMC, Semantic Scholar, and OpenCitations; Crossref provides counts and references only.
Inputs: id (required), limit, sections, citation_mode, +3 more.
article_searchARead-only
Search literature across multiple backends with federated search and deduplication. Response is an envelope: { results, warnings?, errors? } — per-source failures ride in errors (never as items inside results), degradation notes (e.g. PubMed dropping unrecognized query terms, LitSense relevance-floor drops) in warnings. LitSense rows carry the matched sentence plus PubMed-enriched title/metadata when available.
Inputs: query (required), limit, offset, source, +1 more.
batch_getARead-only
Get multiple entities in parallel. Article items return metadata (with inline abstract) by default; for full text pass sections:["fulltext"] on the item — PMC articles then return a small file handle (file_path + outline, never body text inline) while non-PMC articles return an explicit unavailable + reason. Nonexistent PMIDs yield a per-item error and the batch continues. For inline text use article_get with fulltext_output="inline"
Inputs: inputs (required).
biomcp_configureADestructiveIdempotent
Inspect and configure biomcp — unified observability and restricted controllability for every parameter.
What it covers: three optional feature groups (database, analysis_r, analysis_biowasm) whose knobs live in the .biomcp.json project config file (written by this tool, loaded at server startup — a restart applies changes), plus every environment-only parameter (API keys, proxy, security boundaries), which is query-only.
Actions:
status (default; call with {}): per-feature running state (with each feature's settable_keys), config file health, conflicts (e.g. an env var vetoing the file), pending-restart flags, dependency prerequisites, and parameter counts. The full parameter catalog is NOT inlined — use filter for detailed rows: 'file', 'env', a feature id ('analysis_r'), or a dotted-id prefix ('features.database').
set: create/modify file parameters in one atomic batch. Enable/disable is just features..enabled. null removes a key (restores default). Sensitive keys (connection targets, mirrors) require confirm_sensitive=true. Nothing is written when any key is invalid (dry_run=true validates and diffs without writing).
reset: remove a feature section (target: feature id) or specific keys (target: [dotted ids]).
Hard rules: environment parameters are never settable here (the response tells you how to set them in your client's env block); env var values are never displayed (masked: presence + fingerprint only); the file write is refused when the server's working directory is not a project root — the response then carries a paste-ready env block instead.
Examples: {"action":"status"} · {"action":"status","filter":"features.analysis_r"} · {"action":"set","values":{"features.analysis_biowasm.enabled":true}} · {"action":"set","values":{"features.database.enabled":true,"features.database.type":"sqlite","features.database.sqlite_path":["data/geo.db"]}} · {"action":"set","values":{"features.analysis_r.mirror_url":null}} · {"action":"reset","target":"analysis_r"}
Changes apply at server startup: finish dependency prerequisites first (see prerequisites in the response), then restart the client/session once, then re-call this tool with {} to verify running_now.
Inputs: action, filter, target, values, +2 more.
discoverARead-only
Free-text concept resolution - find entities matching a free-text query. Prefers a candidate whose name/symbol exactly matches the query (case-insensitive), ranks those exact matches first across entity types, and demotes numbered susceptibility-locus labels (e.g. "lung cancer susceptibility 1") within each backend's candidates; each result's matched_on field reports why it was picked ("exact" or "first")
Inputs: query (required).
disease_drugsARead-only
Get drugs for a disease via OpenTargets. Results are ranked by max clinical phase, then trial-report count, with salt forms collapsed. Response is an envelope: { results, warnings?, errors? } — OpenTargets failures and unresolved diseases ride in errors[] as { source, message, retryable? } instead of inline items (issue #130)
Inputs: disease_id (required), limit.
disease_getCRead-only
Get detailed disease information by ID
Inputs: disease_id (required), limit, sections.
disease_searchARead-only
Search for diseases by name, phenotype, or keyword. Response is an envelope: { results, warnings?, errors? } (issue #130)
Inputs: query (required), limit, offset.
disease_trialsBRead-only
Get clinical trials for a disease. Response is an envelope: { results, warnings?, errors? } (issue #130)
Inputs: disease_id (required), limit.
drug_getCRead-only
Get detailed drug information by name
Inputs: name (required), limit, sections.
drug_searchARead-only
Search for drugs by name or keyword (chemical identity: names, synonyms, and identifiers in ChEMBL/MyChem). This is NOT a mechanism or target search — MyChem has no mechanism text in this query, so "EGFR inhibitor"-style queries return nothing. For drugs targeting a gene/mechanism, use gene_drugs with the target gene symbol.
Inputs: query (required), limit, offset.
drug_trialsBRead-only
Find clinical trials for a drug
Inputs: drug (required).
ensembl_consequenceARead-only
Predict the functional consequence of a variant on demand via Ensembl VEP — works even for NOVEL variants absent from every database, and for non-human species.
Input forms: HGVS c./p./g. notation ("NM_004333:c.1799T>A", "NM_004333.6:c.1799T>A") or a dbSNP rsID ("rs113488060"). Prefer HGVS over rsIDs when both are available — rsIDs are resolved through dbSNP's coordinate mapping, which can yield less specific consequences. RefSeq accessions may carry a version ("NM_004333.6:c.1799T>A"); when the version is not carried by the current Ensembl release the call is retried against the bare accession and the result flags the reinterpretation in warnings. Per-species rules: the variant must be valid on a transcript of THAT species — RefSeq NM_ accessions work cross-species only when the transcript belongs to it (e.g. "NM_011640:c.100A>G" + species 'mouse' is mouse Trp53; a human NM_ fed to species 'mouse' fails). Ensembl transcript IDs (ENST…, ENSMUST…) are accepted for their own species, and c./p. positions must be in range for the transcript. Invalid input returns the upstream VEP error explaining what could not be mapped. Returns the most severe consequence plus per-transcript effects (impact, codon/amino-acid change, SIFT/PolyPhen where available) and co-located known variants (ClinVar/COSMIC IDs, gnomAD/1000G frequencies when present). For KNOWN human variants, variant_get additionally provides deep pre-computed scores (CADD, REVEL, AlphaMissense, ClinVar stars) — prefer it there.
Inputs: variant (required), limit, species.
ensembl_homologyARead-only
Find orthologues or paralogues of a gene across species (Ensembl Compara) — the cross-species gene mapping source in biomcp.
Returns target stable IDs, gene symbols (when the assembly provides them, e.g. ARAF/RAF1 for BRAF paralogs), species, taxonomy level, and percent identity, sorted by identity. Accepts an HGNC symbol or Ensembl gene ID; scope to one species with target_species ('mouse') or target_taxon (10090). Use it for conservation questions ('is this gene conserved?', 'what is the mouse orthologue of BRAF?').
Inputs: gene (required), type, limit, species, +2 more.
ensembl_lookupARead-only
Resolve a gene in Ensembl terms for ANY Ensembl species (356 available): stable ID (ENSG…), symbol↔ID mapping, versioned identifier, canonical transcript, and coordinates on the current assembly (GRCh38 human, GRCm39 mouse, …). With expand=true, returns all transcripts with translation/protein IDs.
Accepts an HGNC symbol (BRAF) or Ensembl gene ID (ENSG00000157764, versioned or bare — versions are resolved to the current record). Species accepts scientific names or aliases ('homo_sapiens'/'human', 'mus_musculus'/'mouse'). For rich human gene annotation (summary, pathways, drugs, diseases) use gene_get instead — this tool is the identifier/structure authority.
Inputs: gene_or_id (required), expand, species.
ensembl_regionARead-only
Query what lives in a genomic interval on the current assembly (GRCh38 for human): genes/transcripts (stable IDs, symbols, biotypes) and known variants (rsIDs, alleles, consequence terms, clinical significance).
Ideal for locus triage — "what genes and known variants sit in this GWAS hit interval?". Genes/transcripts are always fetched for the FULL requested span, but variation density can reach ~100k variants in dense loci (which times out upstream), so when the span exceeds 50 kb variation is queried only for a 50 kb window centered on the region midpoint — the result echoes effective_variation_window (the coordinates actually queried for variation) and variation_truncated (names the dropped span) when this cap applies. To sweep variation across a large interval, issue adjacent region calls. Keep spans modest (<1 Mb recommended); output is capped at limit with a truncated marker. For sequence text use genbank_get; for entity-level annotation chain IDs into gene_get / variant_get.
Inputs: region (required), limit, species, features.
genbank_genesARead-only
Map a GenBank/RefSeq nucleotide accession to its NCBI Gene IDs (elink nuccore→gene).
The returned gene_ids are NCBI Gene (entrezgene) IDs. gene_get takes an HGNC gene symbol, NOT a numeric entrez id, so an id must first be resolved to its symbol — gene_search result rows carry both symbol and entrez_id, so the id→symbol bridge exists via gene_search. Useful bridge from a sequence record to gene-level annotation.
Inputs: accession (required).
genbank_getARead-only
Fetch a GenBank/RefSeq nucleotide record as text (GenBank flat file or FASTA).
Whole-record fetches are capped at 2,000,000 bp — larger records require a seq_start/seq_stop region (1-based, inclusive, up to 10 Mb span; set strand=2 for a reverse-strand slice where seq_start > seq_stop — the region field echoes the request as given, while NCBI's text shows complement(min..max)). Output: records up to 1,000 characters are returned inline (sequence_text with char_count/truncated). Anything larger defaults to file mode — the FULL text is written to a server-side cache file and the reply carries mode:"file", file_path, char_count, and total_char_count instead of sequence_text (never truncated); Read/Grep the file selectively. The cache is ephemeral (pruned after ~7 days / 500 files) and the path is server-local; re-call to regenerate. sequence_output:"inline" opts back into inline text capped at 200,000 characters — use it only if you cannot read server-side files.
Inputs: accession (required), format, strand, seq_stop, +3 more.
genbank_searchARead-only
Search NCBI nucleotide records (GenBank/RefSeq/INSDC).
Queries may be plain terms, an accession, or NCBI field syntax ("TP53[Gene Name] AND Homo sapiens[Organism]", "BRCA1[Gene Name]"). Results include accession.version, definition, length_bp, organism, and topology — chain accessions into genbank_get or genbank_genes.
Inputs: query (required), limit, offset, organism.
gene_articlesARead-only
Find articles about a gene. Response is an envelope: { results, warnings?, errors? } — per-backend failures ride in errors (never as items inside results), degradation notes in warnings (issue #130)
Inputs: symbol (required).
gene_diseasesARead-only
Get diseases associated with a gene. Requires DISGENET_API_KEY environment variable for DisGeNET data; falls back to OpenTargets gene-disease associations when unavailable.
Inputs: symbol (required), limit.
gene_drugsARead-only
Find drugs targeting a gene. Results are ranked by max clinical phase, then trial-report count, with salt forms collapsed. Response is an envelope: { results, warnings?, errors? } — OpenTargets failures and unresolved genes ride in errors[] as { source, message, retryable? } instead of inline items (issues #130, #179)
Inputs: symbol (required).
gene_enrichARead-only
Perform pathway enrichment analysis for a gene list (Reactome over-representation). Requires at least 3 genes; the list is deduplicated (case-insensitive) before submission. Each pathway reports genes_overlap — the count of unique SUBMITTED genes Reactome matched there, resolved with one follow-up token lookup (Reactome's entity-level match count can exceed the gene count when one gene maps to multiple entities, e.g. a UniProt entry plus Ensembl gene/transcript). The reply also reports how many pathways Reactome found in total alongside the returned page. Typical completion is 15–45 s (the Reactome analysis takes 8–22 s per step across two sequential steps), with a worst case around 105 s including one retry per step; requires reachability of reactome.org, and clients whose request timeout is 60 s may need to raise it for large result follow-ups.
Inputs: genes (required), limit.
gene_getARead-only
Get detailed gene information by symbol. Response section keys mirror the requested section names: sections:["clinical_evidence"] returns sections.clinical_evidence, "protein_atlas" returns sections.protein_atlas, "dosage_sensitivity" returns sections.dosage_sensitivity, "disease_associations" returns sections.disease_associations — the legacy storage keys (civic, hpa, clingen, disgenet) never appear in responses.
Inputs: symbol (required), limit, smart, sections.
gene_searchCRead-only
Search for genes by symbol, name, or keyword
Inputs: query (required), limit, offset, chromosome.
gene_trialsARead-only
Find clinical trials for a gene
Inputs: symbol (required).
geo_getA
Get the full SOFT record details for a GEO entry: series (GSE...), sample (GSM...), or platform (GPL...).
A series detail includes summary, organisms, platform_ids, a sample preview, supplementary file URLs, and cross-references for chaining: sra → sra_get(SRP.../SRR...), pubmed_ids → article_get, samples → geo_get(GSM...), platform_ids → geo_get(GPL...).
GDS accessions (curated DataSets, not served by SOFT) return structured guidance as data — { accession, guidance, see_also: { series: "GSE…", sample: "GSM…" } } — pointing at the underlying series/sample records; this is a successful result, not an error. Malformed accessions still fail at schema validation.
download=true additionally saves a supplementary file to a local temp path and returns its path, size, and source URL — the pick prefers .gz/.csv/.txt, else the first listed supplementary file (subject to max_bytes).
Inputs: accession (required), download, max_bytes.
geo_searchARead-only
Search NCBI GEO (Gene Expression Omnibus) for functional genomics studies — expression microarrays, RNA-seq, and other high-throughput sequencing series.
Queries accept plain terms ("breast cancer RNA-seq", "melanoma single cell") or NCBI field syntax ("GSE183947[Accession]"). Every result carries accession (→ geo_get for full details), bioproject, and pubmed_ids (→ article_get). sra_project (→ sra_get) appears only when NCBI has linked the SRA study to the series — many newer series lack the link — so geo_get(accession) is the reliable route to SRA accessions (its sra field).
Inputs: query (required), limit, offset, organism, +1 more.
gtex_eqtlARead-only
Get significant cis-eQTL associations for a gene in a specific GTEx tissue (GTEx Analysis v10): variant_id, p_value, and NES (effect direction/size), sorted by ascending p-value.
tissue is a required GTEx tissueSiteDetailId (e.g. Whole_Blood, Brain_Cortex).
An empty associations list is a VALID scientific result — the gene was tested but has no significant cis-eQTLs in that tissue (not a lookup failure). In that case the result carries a note listing other tissues where significant eQTLs DO exist (or stating that none exist in any tissue).
Inputs: gene (required), tissue (required), limit.
gtex_expressionARead-only
Get median gene expression across GTEx tissues (GTEx Analysis v10, 54 tissue sites, TPM, sorted highest first).
Accepts an HGNC symbol (TP53) or Ensembl gene ID (ENSG00000141510, versioned or bare). Optionally filter to a single tissue via its tissueSiteDetailId (e.g. Brain_Cortex, Whole_Blood).
Inputs: gene (required), limit, tissue.
journal_metricsARead-only
Journal quality metrics — an open-access alternative to the Clarivate impact factor (IF). Returns the OpenAlex 2-year mean citedness score (an IF-style metric), h-index, i10-index, citation counts, open-access flag (is_oa), and quality-list tiers (e.g. medline, norway-2, jufo-3…). Accepts a journal name, NLM abbreviation (e.g. 'N Engl J Med'), or ISSN (e.g. '0028-0836'; article records carry journal_issn).
Inputs: query (required), limit.
patent_getARead-only
Get patent details by publication number (e.g. "US11027025B2", "EP3904939B1", "US20260240819A1"). Sections: abstract, claims (US fulltext via USPTO Public Search; EP/WO via EPO OPS when configured, otherwise Google Patents pages — see per-section source fields), citations (backward + forward), family, classifications. US claims may fall back to an unofficial Google Patents rendering when PPUBS is slow or unreachable — flagged in _warnings. Section failures are reported per section (inline error entry plus a _warnings note), never by failing the whole call.
Inputs: patent_id (required), limit, sections.
patent_searchARead-only
Search patents worldwide (US, EP, WO, JP, and 100+ authorities). Quote exact multi-word concepts (e.g. "flow cytometry") to avoid off-topic matches. Backend characters: ppubs = USPTO Public Search full-text conceptual search (US only, keyless, relevance-ranked; default US backend) | ops = EPO OPS worldwide bibliographic search over titles/abstracts (needs EPO_OPS_CONSUMER_KEY/EPO_OPS_CONSUMER_SECRET) | uspto_odp = US application metadata, bibliographic only but inventor/CPC/continuity-rich (needs USPTO_API_KEY) | google_patents = worldwide best-effort (often unavailable). Auto mode queries worldwide + ppubs concurrently; if ppubs fails hard it falls back to uspto_odp once (noted in warnings). Per-backend failures and degradation notes are reported out of band in the response errors/warnings arrays — every element of patents is a real patent record and respects limit. Pass source to force a specific backend. Results are ranked by relevance by default (ppubs sort_by). Foundational prior art is auto-discovered via co-citation mining of the top results and returned in seminal_prior_art (surfaces seminal documents whose own vocabulary predates the query concept).
Inputs: query (required), cpc, limit, offset, +7 more.
pdbA
Access the Protein Data Bank (RCSB PDB) for macromolecular structure data.
SEARCH: Provide "query" to find structures (e.g., "kinase inhibitor", "hemoglobin"). GET: Provide "pdb_id" to retrieve entry metadata with optional detail sections. Sections list per-entity records (polymer_entities, ligands, assembly); when some records fail while siblings succeed, each failure appears as an inline {_error} item AND the response carries a top-level "warnings" array naming the section and failed item. DOWNLOAD: Provide "pdb_id" + download=true to save the structure file and get the file path.
Inputs: limit, query, format, offset, +3 more.
preprint_fulltextARead-only
On-demand bioRxiv/medRxiv preprint full text (JATS XML, parsed to sections). Input: a preprint DOI (e.g. "10.1101/2025.03.05.641768"), optional version (default: latest). Downloads are strictly on-demand (never run during search), paced, and cached per (doi, version) — JATS per version is immutable, so repeat calls are served from cache without re-downloading. www.biorxiv.org rate-limits aggressively: on a 429/Cloudflare verdict this tool pauses downloads for >=60s (Retry-After honored) and every call during the cooldown returns an honest unavailable/rate_limited envelope with retry_after_sec — cached copies are still served instantly. The per-article license (cc by, cc by-nc-nd, na, ...) rides on every reply and governs reuse; for corpus-scale text mining use the bioRxiv/medRxiv monthly S3 buckets (see https://www.biorxiv.org/tdm). Output "file" (default): complete text written to a server-side markdown file, reply carries file_path + outline (Read/Grep selectively); "inline": capped text sections (fulltext_max_chars, default 40000).
Inputs: id (required), output, version, fulltext_max_chars.
sra_getARead-only
Get full details for an SRA accession indexed by NCBI: run (SRR/ERR/DRR — instrument, spots, bases, size), experiment (SRX/ERX/DRX — library design), study (SRP/ERP/DRP — experiment list), or sample (SRS/ERS/DRS). NCBI SRA indexes NCBI, European (ENA), and DDBJ submissions alike, so accessions returned by sra_search resolve here regardless of prefix.
Chain from geo_get (sra field) or sra_search (experiment_accession / first_run_accession).
Inputs: accession (required).
sra_searchARead-only
Search NCBI's Sequence Read Archive (SRA) for sequencing experiments and runs.
The query may be free text, an accession (SRP study, SRX experiment, SRR run, SRS sample), or NCBI field syntax ("RNA-SEQ AND Homo sapiens[Organism]"). Results list experiment/study/sample accessions (NCBI S…, European E…, or DDBJ D… prefixes — all resolvable by sra_get), organism, library strategy, run count, and first_run_accession for chaining into sra_get.
Inputs: query (required), limit, offset.
trial_getCRead-only
Get detailed trial information by NCT ID
Inputs: nct_id (required), limit, sections.
trial_searchCRead-only
Search clinical trials by condition, intervention, or keyword
Inputs: query (required), limit, phase, status, +2 more.
variant_getARead-only
Get detailed variant information with optional sections. Core data (id, gene, rsid, significance) is always returned at the top level. Use sections to request additional data. Response section keys mirror the requested section names: sections:["alphagenome_scores"] returns sections.alphagenome_scores — the legacy storage key (alphagenome) never appears in responses.
Inputs: id (required), limit, sections.
variant_oncokbBRead-only
Get OncoKB annotations for a variant in a cancer gene. Requires ONCOKB_TOKEN environment variable.
Inputs: gene (required), protein_change (required).
variant_searchARead-only
Search for variants. Use structured parameters for best results:
rsid: e.g. query="rs113488022"
HGVS: e.g. query="NM_004333.4:c.1799T>A"
Gene filter: e.g. gene="BRAF" with hgvsp="V600E" or consequence="missense"
ClinVar significance: e.g. significance="pathogenic" Do NOT use compound free-text like "BRAF V600E" — use separate gene and hgvsp parameters instead.
Inputs: gene, rsid, hgvsc, hgvsp, +7 more.
variant_trialsARead-only
Find clinical trials for a variant. Response is an envelope: { results, warnings? } — when MyVariant.info ID resolution fails, a warning rides alongside the results (which then come from searching the raw identifier) so a backend failure never masquerades as "no trials exist"; retry may resolve BRAF V600E-style naming (issue #198)
Inputs: variant (required).
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 43 tools
Most tools follow a clear entity+action split (gene_/variant_/drug_/disease_/article_/trial_/geo_/sra_/genbank_/patent_ + search/get), and descriptions actively steer agents apart (e.g. drug_search explicitly says it is not a target search, use gene_drugs; gene_get vs ensembl_lookup is carved out). A few overlaps remain that will cause misselection: five routes into clinical trials (drug_trials, gene_trials, variant_trials, disease_trials, trial_search), preprint_fulltext versus article_get on a preprint DOI, and discover versus the per-entity *_search tools. Batch_get also shadows every get tool.
The dominant convention is consistent snake_case entity_verb (gene_search/gene_get, variant_trials, article_get, geo_search, patent_get), which makes the set predictable. Deviations exist: pdb is a bare noun doing search/get/download in one tool, discover and batch_get are generic verbs without an entity, and journal_metrics/gene_enrich are noun_verb or verb-only rather than entity_action.
43 tools is heavy, but the breadth is genuine (genes, variants, drugs, diseases, trials, literature, patents, PDB, GEO/SRA/GenBank, GTEx, Ensembl, config), and most domains are well-scoped search/get pairs that earn their place. Still, several convenience wrappers (gene_articles, gene_diseases, gene_drugs, gene_trials, batch_get) could be folded into generic search tools, so the surface feels larger than strictly necessary.
Coverage of the bioinformatics domain is broad and lifecycle-complete for most entities: search+get for genes, variants, drugs, diseases, trials, articles, patents, sequences, structures, with rich cross-linking hints (sra→sra_get, geo→sra, genbank_genes→gene_search). Minor gaps: Reactome enrichment is the only pathway tool and has no reverse lookup, there is no protein/UniProt target tool separate from PDB, and no batch/list-upload path except the generic batch_get and gene_enrich gene lists.
Maintenance
Related MCP Connectors
MCP gateway federating 26 biomedical MCP servers behind one endpoint: gnomAD, ClinVar, HPO, VEP.
Auditable MCP server for PubMed, Europe PMC, ClinicalTrials.gov, and bioRxiv/medRxiv queries
MCP server for your apps' tools and custom tools, plus hosted AI agents and approval-gated workflows
MCP server for progressive tool usage at any scale (see https://klavis.ai)
Related MCP Servers
- AlicenseBqualityBmaintenanceAn MCP server that provides standardized access to biomedical knowledge bases and resources, enabling AI systems to retrieve verified information from sources like bioRxiv, EuropePMC, and various protein/gene databases.5233Apache 2.0
- AlicenseBqualityDmaintenanceThis MCP server provides 16 intelligent tools for searching, retrieving, and linking biomedical literature from PubMed and PMC. It enables LLM applications to perform complex queries, batch processing, and cross-database linking.168MIT
- AlicenseNot gradedqualityCmaintenanceUnified MCP server providing AI-agent-ready access to AlphaFold, PubMed, ChEMBL, Ensembl, and 37+ scientific databases.MIT
- AlicenseAqualityAmaintenanceAn MCP server that gives AI assistants access to biological and biomedical RDF databases via SPARQL at the RDF Portal, as well as selected REST APIs (NCBI E-utilities, UniProt, ChEMBL, PDB, Reactome, Rhea, MeSH, and more).23321MIT